fix(desktop): keep active-turn badges through transient relay drops

[wpfleger96]

Problem

Active-turn timer badges on the Agents menu all vanish at once on a transient relay drop (flaky VPN). When the relay drops, liveness frames stop arriving for every agent simultaneously, and the 5s pruneExpired tick deletes every turn whose lastActivityAt is older than REMOVE_AFTER_MS (25s) in one sweep — wiping all badges together. A prior fix addressed frame recovery, not this prune layer.

Fix

All changes are in activeAgentTurnsStore.ts.

• Pause the prune when every tracked turn is simultaneously stale — the "all at once" local signature of a relay drop. pruneExpired early-returns when the max lastActivityAt across all agents' turns is older than FRAME_GAP_PAUSE_MS (20s, below the 25s prune bound). Gating on the max per-turn lastActivityAt rather than a global frame clock is what prevents over-pausing: a single live sibling turn keeps the max fresh, so a genuine multi-agent crash still prunes the dead turn at 25s — no regression. No relay-connection coupling; wake-on-resume is automatic once frames flow again.
• Resurrect a pruned badge on a recovered liveness/acp frame. A turn_liveness/acp_* frame for a turn no longer in the live map recreates it. This runs only for frames that pass the existing per-agent watermark, so replayed stale frames cannot revive a turn.
• Per-turn terminal tombstone (terminalAtByAgent) records when a turn terminally ended (turn_completed/turn_error/agent_panic), and resurrection revives a turn only when the recovered frame is strictly newer than that terminal — a completed turn is never revived. The map is bounded per agent and cleared in resetActiveAgentTurnsStore.

Accepted residual

A lone turn kill -9'd under a healthy relay (it was the only active turn) keeps its badge until its next frame instead of clearing at 25s. This is intrinsic to local-only sensing — that case is locally indistinguishable from a drop — and the badge self-heals the instant any frame arrives. Signed off as the chosen tradeoff to keep badges visible through transient drops.

Notes for review

• The badge render is exercised E2E-only; this PR covers the store logic with the existing node:test unit suite (mock.timers).
• Two prior backstop tests asserted the old contract that a lone silent turn prunes at 25s; they now assert the new contract (dead turn prunes when a live sibling keeps the max fresh; lone silent turn lingers as the documented residual).

[wpfleger96]

Active-turn badge resilience — E2E proof

The badge render is E2E-only, so these screenshots prove the fix on the actual rendered timer badge — not just the unit suite. The spec installs page.clock before navigation, seeds two agents working across channels, then fastForwards 30s with no further frames (firing several real 5s prune ticks past both the 20s pause and 25s remove thresholds) and asserts the badges persist. fastForward (not setFixedTime) is load-bearing: it genuinely fires the prune interval, so shouldPausePrune is actually exercised rather than vacuously skipped.

Healthy multi-agent state (before the drop)

Two agents working across channels, all per-channel Working in #… · <elapsed> timer badges present — Duncan in #general, Paul in #engineering and #general. This is the state Will sees with correct timers before a flaky-VPN relay drop.

Badges survive the all-at-once liveness gap (the resilience proof)

Same view after advancing the clock 30s with no frames. All three badges are still present and their elapsed counters advanced to · 30s. Under the pre-fix code every badge would have vanished at the first prune tick past 25s — this is the exact regression #1120 fixes. The · 30s advance independently confirms the prune ticks fired and the render updated, so shouldPausePrune is what kept the badges alive, not the absence of a tick.